Who Benefits from Religious Attendance?

Heterogeneous Causal Effects on Well-being and Cooperation in New Zealand

Joseph A. Bulbulia

Victoria University of Wellington, New Zealand

New Zealand Attitudes and Values Study in a Nutshell

  • Probability panel (\aprox 1% of electorate; 70–80% annual retention).
  • Postal questionnaires; retention 70–80%
  • Multidisciplinary team > 60 researchers
  • Domains: personality, social attitudes, values, religion, employment, prejudice, religion \dots

Three causal domains for today**

  1. Flourishing – does it enhance multi‑dimensional well‑being (“the good life”)?
  2. Cooperation – does attendance spur prosocial behaviour?
  3. Individual Differences - who responds differently?

Objectives

By the end you should be able to

  1. Specify counterfactual estimands for service attendance
  2. Describe the targeted‑learning pipeline we use to estimate them
  3. Interpret heterogeneous treatment effects (policy trees)

Why Causality & Heterogeneity Matter

  • Multi‑dimensional wellbeing and human cooperation are twin targets of psychological science.
  • Observational links tie monthly religious service attendance to higher wellbeing and prosociality, yet correlation \neq causation.
  • We combine 15 years of NZAVS panel data with semi‑/non‑parametric ML to estimate causal effects and uncover who gains, who loses, and why (Hainmueller, Mummolo, and Xu 2019; Tibshirani et al. 2024).
  • Early evidence points to meaning in life and self‑esteem—rather than social belonging/support—as durable drivers of cooperative behaviour.
  • Sexuality, long framed as the “fundamental problem of cooperation”, also modulates these pathways.
  • Strategic life‑course factors (age, health, resources) shape how attendance translates into meaning, esteem, and cooperation.

Daily survey throughput

Eligibility

  • Responded at Wave 10 (2018/19) baseline
  • Reported service attendance at baseline (Yes/No)
  • Inverse‑probability of censoring weights adjust attrition
  • “Censored” if data collected during national lockdown (Wave 10)
  • Final analytic sample: 46,377 adults.

Our Approach in Plain Language

  1. Track same people over 6 years (n = 46,377)
  2. Compare those who change attendance vs those who don’t
  3. Use AI to find who benefits most
  4. Account for selection bias and dropout
  5. Focus on causal not correlational effects

Note

Technical details in appendix

Context

Sample Demography

Table 1

Demographic statistics by religious denomination (NZAVS wave 10, approx~1.4% of NZ adult population).

Analytic Pipline

Stage Tool Purpose
1. Target Causal estimand definition Binary & shift interventions
2. Learn SuperLearner (semi-parametric ML) Outcome & censoring models
3. Adjust IPC weights + census post-stratification Population inference
4. Validate 10-fold cross-validation Out-of-sample performance
5. Subgroup Causal forests & policy trees Detect heterogeneity

Decoding the Technical Terms

  • Treatment Effect: How much religious attendance helps
  • Heterogeneity: Different people benefit differently
  • Causal Forest: Machine Learning that finds who benefits most
  • QINI: Uplift model performance metric
  • RATE: Relative advantage of targeted treatment
  • CATE: Conditional Average Treatment Effect
  • AIPW: Augmented Inverse Probability Weighting

Simplified Estimands Overview

Estimand Description Example
Binary All or nothing 0 vs 1 service/month
Shift-up Increase unless weekly 0$1, 12, 23, 34 | | Shift-down | Decrease to zero | 40, 30, 20, 1$0

Study 1 | Flourishing (“The Good Life”)

1.1 Binary intervention (3 waves)

Figure 1: ATE: Three-Wave: Binary Intervention

1.2 Shift‑up vs binary (3 waves)

Figure 2: ATE: Three-wave Shift Intervention

1.3 Sensitivity: socialising hours (3 waves)

Figure 3: ATE: Three-wave Soft Intervention

1.4 Duration: 3 vs 6 waves

Figure 4: ATE: Six-wave Soft Intervention GAIN: +1

1.5 Gain vs loss (6 waves)

Figure 5: ATE: Six-wave Soft Intervention LOSS: -1

1.6 Good Life: Compare 6 waves WEEKLY/NULL vs 6 waves ZERO/NULL

Figure 6: CATE RATE

Key findings – Flourishing

  • Meaning & purpose show the most robust positive effect.
  • Binary gains detectable; shift‑based gains weaker but present.
  • Hour‑for‑hour socialising does not explain the effect.
  • Sexual satisfaction follows a similar pattern (consistent with prior literature)
Figure 7: CATE rate

Study 2 | Cooperation

2.1 Binary intervention (3 waves)

Figure 8: ATE: Three-Wave: Binary Intervention

2.2 Shift‑up vs binary (3 waves)

Figure 9: ATE: Three-wave Shift Intervention

2.3 Gain vs loss (3 waves)

Figure 10: ATE: Three-wave Soft Intervention

2.4 Duration: 3 vs 6 waves (gain)

Figure 11: ATE: Six-wave Soft Intervention WEEKLY/ZER0

2.5 Gain vs loss (6 waves)

Figure 12: ATE: Six-wave Soft Intervention LOSS: -1

Key findings – Cooperation

  • Gains in attendance cause small but reliable boosts to donations & neighbourliness.
  • Losses erode cooperation more than gains enhance it (asymmetry).
  • Perceived social support and belonging are not clearly affected
Figure 13: CATE RATE

Study 3a | Individual differences Policy – Flourishing

3.1: Forgiveness

Figure 14: Policy Trees: Forgiveness

Gratitude

Figure 15: Policy Trees: Gratitude

Fatigue

Figure 16: Policy Trees for Hlth Fatigue Outcome

Meaning Sense

Figure 17: Policy Trees: Meaning Sense

Sense of Neighbourhood Community

Figure 18: Policy Trees: Neighbourhood Community

Self Esteem

Figure 19: Policy Trees: Self Esteem

Study 3b | Individual differences Policy – Cooperation

Policy Tree 1: Donations

Figure 20: Policy Trees for log Charity Donate Outcome

3. Policy Trees Summary: Key Moderators Across All Outcomes

Moderators that Matter

  • Age, baseline health, Conscientiousness, Perfectionism
  • Economic strain and hours of unpaid housework
  • Not major moderators: gender or ethnicity

Note

Click tabs below to explore specific policy trees

  1. Signals modest + noisy (NZ context).
  2. Attendance mainly ↑ meaning ↑ prosociality.
  3. Biggest gains for those low in baseline meaning.n.
  • Are meaning and purpose (goals) the pathway by which religion affects cooperation, rather than by the vagaries of perceived belongingness?

Thanks

  • 76,409 individuals who have participated in the New Zealand Attitudes and Values Study since 2009.
  • Templeton Religion Trust
  • University of Auckland
  • Victoria University
  • Georgia State University
  • Don E. Davis
  • Ken Rice
  • Geoff Troughton
  • Chris G. Sibley & other collaborators
  • Grad Students in the EPIC Lap

Thank You!

References

S.1 Heterogeneity Treatment Effect Decision Flow

This following flowchart shows the decision logic:

    START: For each model
             |
             v
    STEP 1: EXCLUSION CHECK
    Is RATE QINI < 0 (stat sig) OR RATE AUTOC < 0 (stat sig)?
             |
        +----+----+
        |         |
       YES       NO
        |         |
        v         v
    EXCLUDED    STEP 2: RATE SELECTION CHECK
    (Stop)      Is RATE QINI > 0 (stat sig) OR RATE AUTOC > 0 (stat sig)?
                         |
                    +----+----+
                    |         |
                   YES       NO
                    |         |
                    v         v
                SELECTED   UNCLEAR
                    |         |
                    v         v
                STEP 3: QINI CURVE ANALYSIS
                (Applied to all non-excluded models)
                    |         |
                    v         v
                Already     Check: Is QINI curve positive
                selected     at any spend level?
                            |
                       +----+----+
                       |         |
                      YES       NO
                       |         |
                       v         v
                   SELECTED   UNCLEAR
                             (final)
Hainmueller, Jens, Jonathan Mummolo, and Yiqing Xu. 2019. “How Much Should We Trust Estimates from Multiplicative Interaction Models? Simple Tools to Improve Empirical Practice.” Political Analysis 27 (2): 163–92. https://doi.org/10.1017/pan.2018.46.
Tibshirani, Julie, Susan Athey, Erik Sverdrup, and Stefan Wager. 2024. Grf: Generalized Random Forests. https://github.com/grf-labs/grf.